If you've searched "parametric memory," you've probably hit two different things: a machine-learning concept, and a product (ours). This page is about the concept — what parametric and non-parametric memory actually mean for large language models — and, at the end, where our product fits.
Parametric memory: knowledge in the weights
Parametric memory is everything a model learned during training, stored implicitly in its parameters — the billions of weights that make up the network. When GPT-class models answer a question from "what they know," they're drawing on parametric memory.
Its strengths and limits both come from the same fact — it's baked in:
- Fast — no lookup; the knowledge is already in the network.
- Fixed — updating it means retraining or fine-tuning.
- Unverifiable — you can't point to where a fact is stored or prove it's unchanged.
- Opaque and forgetful of specifics — it blends everything into weights, so it hallucinates details and can't reliably recall a specific fact you told it last week.
Non-parametric memory: knowledge you can retrieve
Non-parametric memory is knowledge kept outside the model — in a database, a vector store, a knowledge graph, a document index — and pulled in at query time. Retrieval-augmented generation (RAG) is the best-known pattern: fetch relevant text, put it in the context window, let the model reason over it.
Its properties are the mirror image of parametric memory:
- Updatable — add or change a fact instantly; no retraining.
- Inspectable — you can see exactly what was retrieved.
- Attributable — answers can cite their source.
- Persistent — it survives across sessions, unlike the context window.
This is why almost every serious AI agent uses non-parametric memory for anything specific, current, or private: your codebase conventions, a customer's history, last week's decision.
So which do you need?
For an agent that has to remember specific facts across sessions — a coding assistant, a support agent, an ops copilot — you need non-parametric memory. Parametric memory can't be updated per-user or per-project, and it can't prove what it knows. The open questions are then: how do you store it, how do you retrieve the right slice, and can you trust what comes back?
Where Parametric Memory (the product) fits
Yes, our product is named "Parametric Memory," and yes — technically it's a non-parametric system. We keep the name for the brand; here's the honest technical description:
Parametric Memory is a verifiable non-parametric memory substrate for AI agents. Your knowledge lives outside the model and is retrieved on demand, which fixes the updatable / inspectable / persistent problems above. On top of that it adds the two things ordinary non-parametric memory lacks:
- Proof. Every fact is sealed in an RFC 6962 Merkle tree, so your agent can verify its own memory wasn't altered — something a plain vector store can't do.
- Prediction. A Markov layer pre-loads the context an agent is about to need, giving external memory some of the always-there feel of parametric memory, without the fixed, unverifiable downsides.
In other words: the recall persistence and inspectability of non-parametric memory, plus the "it's just there" feel of parametric memory — and a cryptographic guarantee neither side normally offers.
See how verifiable memory works → · Give your agent memory from $5/mo →